Five Thousand Product Photos: Image CDN or Self-Host?

Your store has five thousand product photographs, yet the hosting quote only says unlimited storage. That sounds reassuring until the first catalogue refresh: thumbnails look stale, large images arrive slowly on mobile, and nobody knows which original belongs to which product. The storage allowance answers only one part of the problem.

Choosing an image CDN or self-hosting for ecommerce requires a delivery plan and an editing plan. The right option should make the product clear to buyers while leaving your team able to replace, export and troubleshoot its pictures. Start with a small representative sample before moving the whole library.

Count the versions you actually need

Five thousand originals do not necessarily mean five thousand delivered files. A product might appear as a small search result, a category card, a main photograph and a close-up. Different layouts may need different crops. Count those uses before comparing supplier packages, because your storefront determines the work required from either system.

Create a simple inventory with product identifier, original filename, intended crop and available dimensions. Mark images that must preserve fine texture, transparent backgrounds or small packaging detail. A ceramic bowl photographed against white presents a different compression problem from a label containing instructions. Use both in the trial rather than choosing only attractive, easy examples.

Separate product images from documents and customer uploads. A publicly visible catalogue photograph is not the same thing as a private warranty attachment. Mixing them under one delivery policy makes permissions harder to understand. Keep the ownership and access decision explicit before the implementation team starts copying folders.

Compare three workable arrangements

The simplest self-hosted arrangement stores prepared image sizes alongside the website. Your team generates the required variants when importing products and serves those files directly. This can suit a stable catalogue with a small number of layouts, especially when someone already owns the import process and checks its output.

A second arrangement keeps original assets in storage you control and uses an image service to produce delivery versions. Cloudflare, for example, documents remote-image resizing, format conversion and caching. That is a concrete capability, not a promise that every image service or hosting plan behaves identically. Check the exact product proposed in your quote. Cloudflare image transformations.

A third arrangement moves the library into a managed image platform. That may simplify asset searching and transformation rules, but it adds another account and export process to manage. Ask whether a future supplier can recover originals, product associations and useful metadata without reconstructing them from rendered pages.

Model costs with a normal month and a busy month

Ask for a written cost worksheet using your expected requests, transferred bytes, stored originals and required transformations. Providers can meter different things; a low storage charge does not establish the total monthly bill. Include any charge for moving originals out, and distinguish the image service bill from the website hosting bill.

Use real storefront samples to estimate transferred image weight. Multiply measured page image weight by plausible views of those page types, rather than multiplying every photograph by every visitor. A visitor who opens three category pages and two products sees a different mix from somebody arriving directly on a product advertisement.

Make the busy-month assumptions visible. A catalogue replacement, a campaign and a crawler visiting many products could produce a different workload from routine browsing. Ask the supplier to identify which assumption changes the bill most. You want a range you can review, not a precise-looking number based on unknown traffic.

Decide how replacements become visible

Image replacement is an operational test worth doing before contract approval. Upload a revised picture for one sample product, then inspect its category card, product page and any promotional module. Record whether the same filename is reused and how cached versions are refreshed. A successful upload does not prove every customer will immediately receive the replacement.

One practical approach is to give revised assets a new versioned filename and update the product reference. That makes the selected version explicit, although the team still needs to manage obsolete files and references. If the proposed system relies on a purge command instead, ask who can run it and how completion is checked.

Now imagine a recall of incorrect packaging artwork. Can an editor find every affected product and replace the image without developer intervention? This is a hypothetical acceptance exercise, not a claim about your current catalogue. It exposes whether the media library supports your real correction workflow.

Illustrative concept: one catalogue asset may need several delivery sizes.

Protect visual quality at the sizes buyers use

Evaluate images inside the storefront, not only in the media dashboard. Look at mobile category cards, product zoom and high-density displays. Check whether the crop removes useful edges, whether transparent items acquire an unwanted background, and whether fine detail still helps somebody decide what they are buying.

Ask the developer to show which file the browser actually downloads for each layout. A small card that receives a huge original remains wasteful even if delivery is fast. Conversely, an aggressively reduced image may load quickly but fail to show the feature the customer wants to inspect. Set different rules where the content justifies them.

Include accessibility in the catalogue workflow. Alt text should describe relevant product information where the image communicates it; it should not become a list of repeated sales keywords. Decide which team writes that text and how it travels with the product record during import, replacement and export.

Check product associations as well as appearance. Two nearly identical products may differ in size, connector position or included accessories. An import that assigns the wrong photograph can mislead a buyer even when delivery quality is excellent. Include a few deliberately similar products in the pilot and have someone who knows the catalogue verify the mapping.

Record the review at the product level: identifier, selected original, accepted crops and any exception. This gives editors a useful reference when new stock arrives. Without that record, the next person may repeat an image treatment that looked acceptable for one item but removes an important distinguishing feature from another.

Finally, test a product with no approved image. Agree whether it stays unpublished, uses a clearly neutral placeholder or enters an editorial queue. The image pipeline should make missing work visible rather than quietly copying a neighbouring product's photograph to fill the space.

Make failure and exit part of the comparison

Ask what happens if a transformation request fails or the image account reaches a limit. A fallback needs testing with the proposed platform; an imagined fallback written into a proposal offers little protection. Check the resulting page for broken images, layout shifts and a usable way for the visitor to continue.

Keep originals backed up under an agreed responsibility. A CDN cache is not a substitute for a recoverable asset library. Your acceptance pack should identify where originals live, who controls the account, how exports work and how a replacement developer maps each file back to its product.

Custom image URLs can also create an exit dependency. Request a documented translation between product records and delivery URLs. If a provider-specific transformation string is embedded throughout templates and content, include the effort to change those references when evaluating the convenience offered today.

Run a pilot that answers the purchasing question

Select perhaps twenty representative products as a manageable trial, not as a statistically universal benchmark. Include old and new photographs, varied shapes, transparent backgrounds and products with many variants. Record the current delivery behaviour, implement the proposed arrangement and repeat the same journeys under comparable conditions.

Give an editor a replacement task without coaching. Give a developer an export-and-restore task. Give the commercial owner the normal-month and busy-month worksheets. These three checks cover routine work, continuity and spending; a single speed screenshot cannot cover all of them.

Document what would justify choosing the more complex option. It might be repeated manual resizing, inconsistent crops or expensive delivery under measured traffic. If prepared files solve the present problem cleanly, a full media platform may add administration without enough benefit. If the catalogue changes daily, the balance may be different.

For an ecommerce development brief, send the sample library, current product export, intended layouts and traffic assumptions. Ask for the image workflow to be priced as part of the store, with acceptance checks and account ownership stated. That makes the proposal about operating your catalogue rather than buying a fashionable acronym.